Abstract
Graph convolution networks (GCNs) have emerged as powerful approaches for semi-supervised classification of attributed graph data. In this paper, we introduce a novel objective function designed for training GCNs in an unsupervised learning setting, specifically for clustering purposes. Our proposed loss function is comprised solely of unsupervised components. The first component encompasses the kernel k-means objective function, which captures shared feature information among nodes, while the second component serves as a regularization term, promoting the smoothness of predicted clusters across the entire dataset. By employing this objective function, we preserve the strengths of conventional semi-supervised GCNs while adapting them to the demands of unsupervised clustering tasks. Experimental results, based on standard benchmark datasets, demonstrate that our unsupervised GCN outperforms contemporary state-of-the-art clustering algorithms.
Fadi Dornaika
Ikerbasque Research Professor
Ikerbasque Research Professor with expertise in computer vision, machine learning, and pattern recognition.